U.S. patent application number 10/084587 was filed with the patent office on 2002-12-19 for method for analyzing mass spectra.
Invention is credited to Braginsky, Leonid, Fung, Eric T., Gavin, Edward J., Rich, William E., Wright, George L. JR..
Application Number | 20020193950 10/084587 |
Document ID | / |
Family ID | 22185918 |
Filed Date | 2002-12-19 |
United States Patent
Application |
20020193950 |
Kind Code |
A1 |
Gavin, Edward J. ; et
al. |
December 19, 2002 |
Method for analyzing mass spectra
Abstract
A method that analyzes mass spectra using a digital computer is
disclosed. The method includes entering into a digital computer a
data set obtained from mass spectra from a plurality of samples.
Each sample is, or is to be assigned to a class within a class set
having two or more classes and each class is characterized by a
different biological status. A classification model is then formed.
The classification model discriminates between the classes in the
class set.
Inventors: |
Gavin, Edward J.; (San Jose,
CA) ; Braginsky, Leonid; (Newton, MA) ; Rich,
William E.; (Redwood Shores, CA) ; Fung, Eric T.;
(Cupertino, CA) ; Wright, George L. JR.; (Virginia
Beach, VA) |
Correspondence
Address: |
TOWNSEND AND TOWNSEND AND CREW, LLP
TWO EMBARCADERO CENTER
EIGHTH FLOOR
SAN FRANCISCO
CA
94111-3834
US
|
Family ID: |
22185918 |
Appl. No.: |
10/084587 |
Filed: |
February 25, 2002 |
PCT Filed: |
November 15, 2001 |
PCT NO: |
PCT/US00/44972 |
Current U.S.
Class: |
702/28 |
Current CPC
Class: |
H01J 49/0036
20130101 |
Class at
Publication: |
702/28 |
International
Class: |
G06F 019/00; G01N
031/00 |
Claims
What is claimed is:
1. A method that analyzes mass spectra using a digital computer,
the method comprising: a) entering into a digital computer a data
set obtained from mass spectra from a plurality of samples, wherein
each sample is, or is to be assigned to a class within a class set
comprising two or more classes, each class characterized by a
different biological status, and wherein each mass spectrum
comprises data representing signal strength as a function of
time-of-flight, mass-to-charge ratio, or a value derived from
time-of-flight or mass-to-charge ratio, and is created using a
laser ionization desorption process; and b) forming a
classification model which discriminates between the classes in the
class set, wherein forming comprises analyzing the data set by
executing code that embodies a classification process.
2. The method of claim 1 wherein the mass spectra are selected from
the group consisting of MALDI spectra, surface enhanced laser
desorption/ionization spectra, and electrospray ionization
spectra.
3. The method of claim 1 wherein the class set consists of exactly
two classes.
4. The method of claim 1 wherein the samples comprise biomolecules
selected from the group consisting of polypeptides and nucleic
acids.
5. The method of claim 1 wherein the samples are derived from a
eukaryote, a prokaryote or a virus.
6. The method of claim 1 wherein the different biological statuses
comprise a normal status and a pathological status.
7. The method of claim 1 where the different biological statuses
comprise un-diseased, low grade cancer and high grade cancer.
8. The method of claim 1 wherein the different biological statuses
comprise a drug treated state and a non-drug treated state.
9. The method of claim 1 wherein the different biological statuses
comprise a drug-responder state and a drug-non-responder state.
10. The method of claim 1 wherein the different biological statuses
comprise a toxic state and a non-toxic state.
11. The method of claim 10 wherein the toxic state results from
exposure to a drug.
12. The method of claim 1 wherein the data set is a known data set,
and each sample is assigned to one of the classes before the data
set is entered into the digital computer.
13. The method of claim 1 wherein forming the classification model
comprises using pre-existing marker data to form the classification
model.
14. The method of claim 1 wherein the data set is formed by:
detecting signals in the mass spectra, each mass spectrum
comprising data representing signal strength as a function of
mass-to-charge ratio; clustering the signals having similar
mass-to-charge ratios into signal clusters; selecting signal
clusters having at least a predetermined number of signals with
signal intensities above a predetermined value; identifying the
mass-to-charge ratios corresponding to the selected signal
clusters; and forming the data set using signal intensities at the
identified mass-to-charge ratios.
15. The method of claim 1 wherein forming the classification model
comprises at least one of identifying features that discriminate
between the different biological statuses, and leaning.
16. The method of claim 1 wherein the classification process
comprises a neural network analysis.
17. The method of claim 1 further comprising: c) interrogating the
classification model to determine if one or more features
discriminate between the different biological statuses.
18. The method of claim 1 further comprising: c) repeating a) and
b) using a larger plurality of samples.
19. The method of claim 1 wherein the classification process is a
cluster analysis.
20. The method of claim 1 further comprising forming the data set,
wherein forming the data set comprises obtaining raw data from the
mass spectra and then preprocessing the raw mass spectra data to
form the data set.
21. The method of claim 1 wherein the different classes are
selected from exposure to a drug, exposure to one of a class of
drugs and lack of exposure to a drug or one of a class of
drugs.
22. The method of claim 1 wherein the each mass spectrum comprises
data representing signal strength as a function mass-to-charge
ratio or a value derived from mass-to-charge ratio.
23. A method for classifying an unknown sample into a class
characterized by a biological status using a digital computer, the
method comprising: a) entering data obtained from a mass spectrum
of the unknown sample into a digital computer; and b) processing
the mass spectrum data using the classification model formed by the
method of claim 1 to classify the unknown sample in a class
characterized by a biological status.
23. The method of claim 23 wherein the class is characterized by a
disease status.
24. The method of claim 23 wherein the different biological
statuses comprise un-diseased, low grade cancer and high grade
cancer.
25. The method of claim 23 wherein the class is characterized by
exposure to a drug of one of a class of drugs.
26. The method of claim 23 wherein the class is characterized by
response to a drug.
27. The method of claim 23 wherein the class is characterized by a
toxicity status.
28. A method for estimating the likelihood that an unknown sample
is accurately classified as belonging to a class characterized by a
biological status using a digital computer, the method comprising:
a) entering data obtained from a mass spectrum of the unknown
sample into a digital computer; and b) processing the mass spectrum
data using the classification model formed by the method of claim 1
to estimate the likelihood that the unknown sample is accurately
classified into a class characterized by a biological status.
29. A computer readable medium comprising: a) code for entering
data obtained from a mass spectrum of an unknown sample into a
digital computer; and b) code for processing the mass spectrum data
using the classification model formed by the method of claim 1 to
classify the unknown sample in a class characterized by a
biological status.
30. A system comprising: a gas phase ion spectrometer; a digital
computer adapted to process data from the gas phase ion
spectrometer; and the computer readable medium of claim 29 in
operative association with the digital computer.
31. The system of claim 30 wherein the gas phase ion spectrometer
is adapted to perform a laser desorption ionization process.
32. A computer readable medium comprising: a) code for entering
data obtained from a mass spectrum of an unknown sample into a
digital computer; and b) code for processing the mass spectrum data
using the classification model formed by the method of claim 1 to
estimate the likelihood that the unknown sample is accurately
classified into a class characterized by a biological status.
33. A system comprising: a gas phase ion spectrometer; a digital
computer adapted to process data from the gas phase ion
spectrometer; and the computer readable medium of claim 32 in
operative association with the digital computer.
34. The system of claim 33 wherein the gas phase ion spectrometer
is adapted to perform a laser desorption ionization process.
35. A computer readable medium comprising: a) code for entering
data derived from mass spectra from a plurality of samples, wherein
each sample is, or is to be assigned to a class within a class set
of two or more classes, each class characterized by a different
biological status, and wherein each mass spectrum comprises data
representing signal strength as a function of time-of-flight,
mass-to-charge ratio or a value derived from mass-to-charge ratio
or time-of-flight, and is created using a laser desorption
ionization process; and b) code for forming a classification model
using a classification process, wherein the classification model
discriminates between the classes in the class set.
36. The computer readable medium of claim 35 wherein the
classification process comprises a neural network analysis.
37. A system comprising: a gas phase ion spectrometer; a digital
computer adapted to process data from the gas phase ion
spectrometer; and the computer readable medium of claim 35 in
operative association with the digital computer.
38. The system of claim 37 wherein the gas phase ion spectrometer
is adapted to perform a laser desorption ionization process.
Description
RELATED APPLICATIONS
[0001] This application claims the benefit of U.S. Provisional
Patent Application Nos. 60/249,835 filed Nov. 16, 2000 and
60/254,746 filed Dec. 11, 2000. These U.S. Provisional Patent
Applications are herein incorporated by reference in their entirety
for all purposes.
FIELD OF THE INVENTION
[0002] Embodiments of the invention relate to methods for analyzing
mass spectra.
BACKGROUND OF THE INVENTION
[0003] Recent advances in genomics research have led to the
identification of numerous genes associated with various diseases.
However, while genomics research can identify genes associated with
a genetic predisposition to disease, there is still a need to
characterize and identify markers such as proteins. A "marker"
typically refers to a polypeptide or some other molecule that
differentiates one biological status from another. Proteins and
other markers are important factors in disease states. For example,
proteins can vary in association with changes in biological states
such as disease. They can also signal cellular responses to
disease, toxicity, or other stimuli. When disease strikes, some
proteins become dormant, while others become active. Prostate
Specific Antigen (PSA), for example, is a circulating serum protein
that, when elevated, correlates with prostate cancer. If the
changes in protein levels could be rapidly detected, physicians
could diagnose diseases early and improve treatments.
[0004] Identifying novel markers is one of the earliest and most
difficult steps in the diagnostics and drug discovery processes.
One way to discover if substances are markers for a disease is by
determining if they are "differentially expressed" in biological
samples from patients exhibiting the disease as compared to samples
from patients not having the disease. For example, FIG. 1(a) shows
one graph 100 of a plurality of overlaid mass spectra of samples
from a group of 18 diseased patients.
[0005] The diseased patients could have, for example, prostate
cancer. Another graph 102 is shown in FIG. 1(b) and illustrates a
plurality of overlaid mass spectra of samples from a group of 18
normal patients. In each of the graphs 100, 102, signal intensity
is plotted as a function of mass-to-charge ratio. The intensities
of the signals shown in the graphs 100, 102 are proportional to the
concentrations of markers having a molecular weight related to the
mass-to-charge ratio A in the samples. As shown in the graphs 100,
102, at the mass-to-charge ratio A, a number of signals are present
in both pluralities of mass spectra. The signals include peaks that
represent potential markers having molecular weights related to the
mass-to-charge ratio A.
[0006] When the signals in the graphs 100,102 are viewed
collectively, it is apparent that the average intensity of the
signals at the mass-to-charge ratio A is higher in the samples from
diseased patients than the samples from the normal patients. The
marker at the mass-to-charge ratio A is said to be "differentially
expressed" in diseased patients, because the concentration of this
marker is, on average, greater in samples from diseased patients
than in samples from normal patients.
[0007] In view of the data shown in FIGS. 1(a) and 1(b), it can be
generally concluded that the samples from diseased patients have a
greater concentration of the marker with the mass-to-charge ratio A
than the samples from normal patients. Since the concentration of
the marker is generally greater in samples from diseased patients
than in the normal samples, the marker can also be characterized as
being "up-regulated" for the disease. If the concentration of the
marker was generally less in the samples from diseased patients
than in the samples from normal patients, the protein could be
characterized as being "down-regulated".
[0008] Once markers are discovered, they can be used as diagnostic
tools. For example, with reference to the example described above,
an unknown sample from a test patient may be analyzed using a mass
spectrometer and a mass spectrum can be generated. The mass
spectrum can be analyzed and the intensity of a signal at the
mass-to-charge ratio A can be determined in the test patient's mass
spectrum. The signal intensity can be compared to the average
signal intensities at the mass-to-charge ratio A for diseased
patients and normal patients. A prediction can then be made as to
whether the unknown sample indicates that the test patient has or
will develop cancer. For example, if the signal intensity at the
mass-to-charge ratio A in the unknown sample is much closer to the
average signal intensity at the mass-to-charge ratio A for the
diseased patient spectra than for the normal patient spectra, then
a prediction can be made that the test patient is more likely than
not to develop or have the disease.
[0009] While the described differential expression analysis is
useful, many improvements could be made. For instance, analyzing
the amount of a single marker such as PSA in a patient's biological
sample is many times not sufficiently reliable to monitor disease
processes. PSA is considered to be one of the best prostate cancer
markers presently available. However, it does not always correctly
differentiate benign from malignant prostate disease. While the
concentration of a marker such as PSA in a biological sample
provides some ability to predict whether a test patient has a
disease, an analytical method with a greater degree of reliability
is desirable.
[0010] Also, when a large number of mass spectra of a large number
of biological samples are analyzed, it is not readily apparent
which signals represent markers that might differentiate between a
diseased state and a non-diseased state. A typical mass spectrum of
a biological sample has numerous potential marker signals (e.g.,
greater than 200) and a significant amount of noise. This can make
the identification of potentially significant signals and the
identification of average signal differentials difficult.
Consequently, it is difficult to identify and quantify potential
markers. Unless the potential markers exhibit strong up-regulation
or strong down-regulation, the average signal differential between
samples from diseased patients and samples from normal patients may
not be easily discernable. For example, it is often difficult to
visually determine that a cluster of signals at a given mass value
in one group of mass spectra has higher or lower average signal
intensity than a cluster of signals from another group of mass
spectra. In addition, many potentially significant signals may have
low intensity values. The noise in the spectra may obscure many of
these potentially significant signals. The signals may go
undiscovered and may be inadvertently omitted from a differential
expression analysis.
[0011] It would be desirable to have better ways to analyze mass
spectra. For example, it would be desirable to provide for a more
accurate method for discovering potentially useful markers. It
would also be desirable to provide an improved classification model
that can be used to predict whether an unknown sample is associated
or is not associated with a particular biological status.
[0012] Embodiments of the invention address these and other
problems.
SUMMARY OF THE INVENTION
[0013] Embodiments of the invention relate to methods for analyzing
mass spectra. In embodiments of the invention, a digital computer
forms a classification model that can be used to differentiate
classes of samples associated with different biological statuses.
The classification model can be used as a diagnostic tool for
prediction. It may also be used to identify potential markers
associated with a biological status. In addition, the
classification model can be formed using a process such as, for
example, a neural network analysis.
[0014] One embodiment of the invention is directed to a method that
analyzes mass spectra using a digital computer. The method
comprises: entering into a digital computer a data set obtained
from mass spectra from a plurality of samples, wherein each sample
is, or is to be assigned to a class within a class set comprising
two or more classes, each class characterized by a different
biological status, and wherein each mass spectrum comprises data
representing signal strength as a function of mass-to-charge ratio
or a value derived from mass-to-charge ratio, and is formed using a
laser desorption ionization process; and b) forming a
classification model which discriminates between the classes in the
class set, wherein forming comprises analyzing the data set by
executing code that embodies a classification process.
[0015] Another embodiment of the invention is directed to a method
that analyzes mass spectra using a digital computer. The method
comprises: a) entering into a digital computer a data set obtained
from mass spectra from a plurality of samples, wherein each sample
is, or is to be assigned to a class within a class set comprising
two or more classes, each class characterized by a different
biological status, and wherein each mass spectrum comprises data
representing signal strength as a function of time-of-flight or a
value derived from time-of-flight, and is formed using a laser
desorption ionization process; and b) forming a classification
model which discriminates between the classes in the class set,
wherein forming comprises analyzing the data set by executing code
embodying a classification process.
[0016] Another embodiment is directed to a computer readable
medium. The computer readable medium comprises: a) code for
entering data derived from mass spectra from a plurality of
samples, wherein each sample is, or is to be assigned to a class
within a class set of two or more classes, each class characterized
by a different biological status, and wherein each mass spectrum
comprises data representing signal strength as a function of
time-of-flight or a value derived from time-of-flight, or
mass-to-charge ratio or a value derived from mass-to-charge ratio,
and is formed using a laser desorption ionization process; and b)
code for forming a classification model using a classification
process process, wherein the classification model discriminates
between the classes in the class set.
[0017] Another embodiment of the invention is directed to a method
for classifying an unknown sample into a class characterized by a
biological status using a digital computer. The method comprises:
a) entering data obtained from a mass spectrum of the unknown
sample into a digital computer; and b) processing the mass spectrum
data using a classification model to classify the unknown sample in
a class characterized by a biological status. The classification
model may be formed using, for example, a neural network
analysis.
[0018] Another embodiment of the invention is directed to a method
for estimating the likelihood that an unknown sample is accurately
classified as belonging to a class characterized by a biological
status using a digital computer. The method comprises: a) entering
data obtained from a mass spectrum of the unknown sample into a
digital computer; and b) processing the mass spectrum data using a
classification model to estimate the likelihood that the unknown
sample is accurately classified into a class characterized by a
biological status. The classification model may be formed using a
classification process, and is formed using a data set obtained
from mass spectra of samples assigned to two or more classes with
different biological statuses.
[0019] In embodiments of the invention, the mass spectra being
analyzed may be pre-existing mass spectra which, for example, may
have been created well before the classification model is formed.
Alternatively, the mass spectra data may have been created
substantially contemporaneously with the formation of the
classification model.
[0020] These and other embodiments of the invention are described
with reference to the Figures and the Detailed Description.
BRIEF DESCRIPTION OF THE DRAWINGS
[0021] FIG. 1(A) shows overlaid mass spectra for samples from
diseased patients.
[0022] FIG. 1(B) shows overlaid mass spectra for samples from
normal patients.
[0023] FIG. 2 illustrates a flowchart of a method for creating mass
spectra according to an embodiment of the invention.
[0024] FIG. 3 shows a graph of log normalized intensity as a
function of identified peak clusters. The signal intensities from
mass spectra from two different groups of samples are shown in the
graph.
[0025] FIG. 4 shows a flowchart illustrating some preferred mass
spectra preprocessing procedures according to an embodiment of the
invention.
[0026] FIG. 5 shows a flowchart illustrating some preferred mass
spectra preprocessing procedures and classification model formation
procedures according to an embodiment of the invention.
[0027] FIG. 6 shows a block diagram of a system according to an
embodiment of the invention.
[0028] FIG. 7 shows a classification and regression tree according
to an embodiment of the invention.
[0029] FIG. 8 shows a table showing the variable importance of
different predictor variables.
[0030] FIG. 9 shows gel views obtained from different samples from
cancer patients and normal patients.
[0031] FIG. 10 show spectral views obtained from different samples
from cancer and normal patients.
DETAILED DESCRIPTION
[0032] In embodiments of the invention, a data set obtained from
mass spectra is entered into a digital computer to form a
classification model. The mass spectra are preferably obtained from
biological samples having known characteristics. In preferred
embodiments, the data set used to form the classification model is
characterized as a "known" data set, because the biological
statuses associated with the biological samples are known before
the data set is used to form the classification model. In
comparison, an "unknown" data set includes data that is obtained
from mass spectra of samples where it is unclear if the samples are
associated with the biological statuses which are discriminated by
the classification model when the mass spectra are formed. Unknown
data may be derived from a biological sample from a test patient
who is to be diagnosed using the classification model. In some
environments, the known data set is referred to as "training
data".
[0033] For purposes of illustration, many of the examples described
below refer to using a known data set to form a classification
model. However, in some embodiments of the invention, the data set
used to form the classification model may be an unknown data set.
For example, in a cluster analysis, mass spectra of unknown
biological samples may be grouped together if they have similar
patterns. Samples corresponding to each group may be analyzed to
see if they have a biological status in common. If so, then the
samples in the group may be assigned to a class associated with the
biological status. For example, after forming a group of mass
spectra having common patterns, it may be determined that all
spectra in the group were obtained from biological samples that
were all exposed to radiation. The samples in the group may then be
assigned to a class that is associated with the status "radiation
exposed". Samples in other groupings can be assigned to classes
characterized by other biological statuses common to the samples in
the respective groupings. A classification model can thus be formed
and unknown spectra may be classified using the formed
classification model.
[0034] In embodiments of the invention, each sample used is, or is
to be assigned to a class of a set of two or more classes, and each
class is characterize by a different biological status. For
example, a first class of samples may be associated with a
biological status such as a diseased state. A second class of mass
spectra of samples may be associated with a biological status such
as a non-diseased state. The samples in the first and second
classes may form the class set. The mass spectra from each of the
respective classes can contain data that differentiates the first
and the second classes.
[0035] In embodiments of the invention, each mass spectrum in the
analyzed mass spectra could comprise signal strength data as a
function of time-of-flight, a value derived from time-of-flight
(e.g. mass-to-charge ratio, molecular weight, etc.), mass-to-charge
ratio, or a value derived from mass-to-charge ratio (e.g.,
molecular weight). As known by those of ordinary skill in the art,
mass-to-charge ratio values obtained from a time-of-flight mass
spectrometer are derived from time-of-flight values. Mass-to-charge
ratios may be obtained in other ways. For example, instead of using
a time-of-flight mass spectrometer to determine mass-to-charge
ratios, mass spectrometers using quadrupole analyzers and magnetic
mass analyzers can be used to determine mass-to-charge ratios.
[0036] In preferred embodiments, each mass spectrum comprises
signal strength data as a function of mass-to-charge ratio. In a
typical spectral view-type mass spectrum, the signal strength data
may be in the form of "peaks" on a graph of signal intensity as a
function of mass-to-charge ratio. Each peak may have a base and an
apex, where peak width narrows from the base to the apex. The
mass-to-charge ratio generally associated with the peak corresponds
to the apex of the peak. The intensity of the peak is also
generally associated with the apex of the peak.
[0037] Generally, the mass-to-charge ratio relates to the molecular
weight of a potential marker. For example, if a potential marker
has a charge of +1, then the mass-to-charge ratio is equal to the
molecular weight of the potential marker represented by the signal.
Thus, while some mass spectra plots may show signal intensity as a
function of molecular weight, the molecular weight parameter is in
fact derived from mass-to-charge ratios.
[0038] While many specific embodiments of the invention discussed
herein refer to the use of mass-to-charge ratios, it is understood
that time-of-flight values, or other values derived from
time-of-flight values, may be used in place of mass-to-charge ratio
values in any of the specifically discussed exemplary
embodiments.
[0039] Although each mass spectrum in the analyzed mass spectra can
comprise signal strength data as a function of time of flight, the
use of mass spectra having signal strength data as a function of
mass-to-charge ratio is generally preferred. Time-of-flight values
for ions are machine dependent, whereas mass-to-charge ratio values
are machine independent. For example, in a time-of-flight mass
spectrometry process, the time-of-flight values obtained for ions
can depend on the length of the free flight tube in the particular
mass spectrometer used. Different mass spectrometers with different
free flight tube lengths can produce different time-of-flight
values for the same ion. This is not the case for mass-to-charge
ratios, since a mass-to-charge ratio is simply the ratio of the
mass of an ion to the charge of the ion. Classification models
created using mass-to-charge ratio values can also be independent
of the particular mass spectrometer used to create them.
[0040] The data set may comprise any suitable data and may be
entered automatically or manually into a digital computer. The data
may be raw or preprocessed before being processed by the
classification process run on the digital computer. For example,
the raw intensities of signals at predetermined mass-to-charge
ratios in the mass spectra may be used as the data set.
Alternatively, the raw data may be preprocessed before the
classification model is formed. For example, in some embodiments,
the log values of the intensities (e.g., base 2) of the signals in
the mass spectra may be used to form the data set.
[0041] The data set is entered into the digital computer. Computer
code that embodies a classification process uses the data set to
form a classification model. Exemplary classification processes
include hierarchical classification processes such as a
classification and regression tree process, multivariate
statistical analyses such as a cluster analysis, and non-linear
processes such as a neural network analysis. In preferred
embodiments, the data set is processed using a classification and
regression tree process to produce a classification model such as a
classification and regression tree. These and other classification
processes and classification models are described in greater detail
below.
[0042] The created classification model may be predictive or
descriptive. For example, the model can be used to predict whether
an unknown test biological sample is or is not associated with a
particular biological status. Alternatively or additionally, the
classification model may be interrogated to identify features in
the data that differentiate the biological status(s) being
analyzed. A feature includes any aspect of the mass spectra data
that can differentiate the particular classes being analyzed.
Suitable features that can be identified include, but are not
limited to, signal intensities or signal intensity ranges at one or
more mass-to-charge ratios, signal shapes (e.g., peak shapes),
signal areas (e.g., peak areas), signal widths (e.g., peak widths
such as at the bottom of a peak), the number of signals in each
mass spectrum, etc. In a typical example, the classification model
may indicate that a feature such as a particular signal intensity
at a given mass-to-charge ratio differentiates diseased samples
from non-diseased samples. In yet another example, the
classification model may indicate that a combination of features
differentiates diseased samples from non-diseased samples. For
example, signal intensity ranges for two or more signals at
different mass-to-charge ratios may differentiate a diseased state
from a non-diseased state.
[0043] In another example, a suitable feature that may be
identified as differentiating the different sample classes may be
the frequency that signals occur at a particular mass-to-charge
ratio within a class. For example, for a diseased class having 100
samples and a normal class having 100 samples, a signal of
intensity Y at a mass-to-charge ratio X may be present in the mass
spectra of 90 diseased class samples, but may be present in only in
10 samples from the normal class samples. Even though the average
intensity of the signals is the same in both the diseased class and
the normal class (i.e., an average intensity of Y), the higher
number of occurrences of the signal in the cancer patient class
indicates that the feature differentiates the diseased class from
the normal class. A frequency feature such as this can be
identified using the classification model.
[0044] Any suitable biological samples may be used in embodiments
of the invention. Biological samples include tissue (e.g., from
biopsies), blood, serum, plasma, nipple aspirate, urine, tears,
saliva, cells, soft and hard tissues, organs, semen, feces, urine,
and the like. The biological samples may be obtained from any
suitable organism including eukaryotic, prokaryotic, or viral
organisms.
[0045] The biological samples may include biological molecules
including macromolecules such as polypeptides, proteins, nucleic
acids, enzymes, DNA, RNA, polynucleotides, oligonucleotides,
nucleic acids, carbohydrates, oligosaccharides, polysaccharides;
fragments of biological macromolecules set forth above, such as
nucleic acid fragments, peptide fragments, and protein fragments;
complexes of biological macromolecules set forth above, such as
nucleic acid complexes, protein-DNA complexes, receptor-ligand
complexes, enzyme-substrate, enzyme inhibitors, peptide complexes,
protein complexes, carbohydrate complexes, and polysaccharide
complexes; small biological molecules such as amino acids,
nucleotides, nucleosides, sugars, steroids, lipids, metal ions,
drugs, hormones, amides, amines, carboxylic acids, vitamins and
coenzymes, alcohols, aldehydes, ketones, fatty acids, porphyrins,
carotenoids, plant growth regulators, phosphate esters and
nucleoside diphospho-sugars, synthetic small molecules such as
pharmaceutically or therapeutically effective agents, monomers,
peptide analogs, steroid analogs, inhibitors, mutagens,
carcinogens, antimitotic drugs, antibiotics, ionophores,
antimetabolites, amino acid analogs, antibacterial agents,
transport inhibitors, surface-active agents (surfactants),
mitochondrial and chloroplast function inhibitors, electron donors,
carriers and acceptors, synthetic substrates for proteases,
substrates for phosphatases, substrates for esterases and lipases
and protein modification reagents; and synthetic polymers,
oligomers, and copolymers. Any suitable mixture or combination of
the substances specifically recited above may also be included in
the biological samples.
[0046] As noted above, the biological samples from which the data
set is created are assigned to a class in a set of two or more
classes. Each class is characterized by a different biological
status. Preferably, there are only two classes and two biological
statuses; one for each of the two classes. For example, one class
may have a biological status such as a diseased state while the
other biological status may have a status such as a non-diseased
state.
[0047] As used herein, "biological status" of a sample refers to
any characterizing feature of a biological state of the sample or
the organism or source from which the sample is derived. The
feature can be a biological trait such as a genotypic trait or a
phenotypic trait. The feature can be a physiological or disease
trait, such as the presence or absence of a particular disease,
including infectious disease. The feature also can be a condition
(environmental, social, psychological, time-dependent, etc.) to
which the sample has been exposed.
[0048] Genotypic traits can include the presence or absence of a
particular gene or polymorphic form of a gene, or combination of
genes. Genetic traits may be manifested as phenotypic traits or
exist as susceptibilities to their manifestation, such as a
susceptibility to a particular disease (e.g., a propensity for
certain types of cancer or heart disease).
[0049] Phenotypic traits include, for example, appearance,
physiological traits, physical traits, neurological conditions,
psychiatric conditions, response traits, e.g., or response or lack
of response to a particular drug. Phenotypic traits can include the
presence of absence of so-called "normal" or "pathological" traits,
including disease traits. Another status is the presence or absence
of a particular disease. A status also can be the status of
belonging to a particular person or group such as different
individuals, different families, different age states, different
species, and different tissue types.
[0050] In some embodiments, the biological statuses may be, for
example, one or more of the following in any suitable combination:
a diseased state, a normal status, a pathological status, a drug
state, a non-drug state, a drug responder state, a non-drug
responder state, and a benign state. A drug state may include a
state where patient who has taken a drug, while a non-drug state
may include a state where a patient has not taken a drug. A drug
responder state is a state of a biological sample in response to
the use of a drug. Specific examples of disease states include,
e.g., cancer, heart disease, autoimmune disease, viral infection,
Alzheimer's disease and diabetes. More specific cancer statuses
include, e.g., prostate cancer, bladder cancer, breast cancer,
colon cancer, and ovary cancer. Biological statuses may also
include beginning states, intermediate states, and terminal states.
For example, different biological statuses may include the
beginning state, the intermediate state, and the terminal state of
a disease such as cancer.
[0051] Other statuses may be associated with different environments
to which different classes of samples are subjected. Illustrative
environments include one or more conditions such as treatment by
exposure to heat, electromagnetic radiation, exercise, diet,
geographic location, etc. For example, a class of biological
samples (e.g., all blood samples) may be from a group of patients
who have been exposed to radiation and another class of biological
samples may be from a group of patients who have not been exposed
to radiation. The radiation source may be an intended radiation
source such as an x-ray machine or may be an unintended radiation
source such as a cellular phone. In another example, one group of
persons may have been on a particular diet of food, while another
group may have been on a different diet.
[0052] In other embodiments of the invention, the different
biological statuses may correspond to samples that are associated
with respectively different drugs or drug types. In an illustrative
example, mass spectra of samples from persons who were treated with
a drug of known effect are created. The mass spectra associated
with the drug of known effect may represent drugs of the same type
as the drug of known effect. For instance, the mass spectra
associated with drugs of known effect may represent drugs with the
same or similar characteristics, structure, or the same basic
effect as the drug of known effect. Many different analgesic
compounds, for example, may all provide pain relief to a person.
The drug of known effect and drugs of the same or similar type
might all regulate the same biochemical pathway in a person to
produce the same effect on a person. Characteristics of the
biological pathway (e.g., up- or down-regulated proteins) may be
reflected in the mass spectra.
[0053] A classification model can be created using the mass spectra
associated with the drug of known effect and mass spectra
associated with different drugs, different drug types, or no drug
at all. Once the classification model is created, a mass spectrum
can then be created for a candidate sample associated with a
candidate drug of unknown effect. Using the classification model,
the mass spectrum associated with the candidate sample is
classified. The classification model can determine if the candidate
sample is associated with the drug of known effect or another drug
of a different type. If, for example, the classification model
classifies the candidate sample as being associated with the drug
of known effect, then the candidate drug is likely to have the same
effect on a person as the drug of known effect. Accordingly,
embodiments of the invention can be used, among other things, to
discover and/or characterize drugs.
[0054] I. Obtaining Mass Spectra
[0055] The mass spectra may be obtained by any suitable process.
For example, the mass spectra may be retrieved (e.g., downloaded)
from a local or remote server computer having access to one or more
databases of mass spectra. The databases may contain libraries of
mass spectra of different biological samples associated with
different biological statuses. Alternatively, the mass spectra may
be generated from the biological samples. Regardless of how they
are obtained, the mass spectra and the samples used to create the
classification model are preferably processed under similar
conditions to ensure that any changes in the spectra are due to the
samples themselves, and not differences in processing. The mass
spectra might be created specifically with a particular
classification process in mind, or might be created without
reference to a particular classification process used on the
data.
[0056] In embodiments of the invention, a gas phase ion
spectrometer mass may be used to create mass spectra. A "gas phase
ion spectrometer" refers to an apparatus that measures a parameter
that can be translated into mass-to-charge ratios of ions formed
when a sample is ionized into the gas phase. This includes, e.g.,
mass spectrometers, ion mobility spectrometers, or total ion
current measuring devices.
[0057] The mass spectrometer may use any suitable ionization
technique. The ionization techniques may include for example, an
electron ionization, fast atom/ion bombardment, matrix-assisted
laser desorption/ionization (MALDI), surface enhanced laser
desorption/ionization (SELDI), or electrospray ionization.
[0058] In some embodiments, an ion mobility spectrometer can be
used to detect and characterize a marker. The principle of ion
mobility spectrometry is based on the different mobility of ions.
Specifically, ions of a sample produced by ionization move at
different rates due to their difference in, e.g., mass, charge, or
shape, through a tube under the influence of an electric field. The
ions (typically in the form of a current) are registered at a
detector and the output of the detector can then be used to
identify a marker or other substances in the sample. One advantage
of ion mobility spectrometry is that it can be performed at
atmospheric pressure.
[0059] In preferred embodiments, a laser desorption time-of-flight
mass spectrometer is used to create the mass spectra. Laser
desorption spectrometry is especially suitable for analyzing high
molecular weight substances such as proteins. For example, the
practical mass range for a MALDI or a surface enhanced laser
desorption/ionization process can be up to 300,000 daltons or more.
Moreover, laser desorption processes can be used to analyze complex
mixtures and have high sensitivity. In addition, the likelihood of
protein fragmentation is lower in a laser desorption process such
as a MALDI or a surface enhanced laser desorption/ionization
process than in many other mass spectrometry processes. Thus, laser
desorption processes can be used to accurately characterize and
quantify high molecular weight substances such as proteins.
[0060] In a typical process for creating a mass spectrum, a probe
with a marker is introduced into an inlet system of the mass
spectrometer. The marker is then ionized. After the marker ions are
generated, the generated ions are collected by an ion optic
assembly, and then a mass analyzer disperses and analyzes the
passing ions. The ions exiting the mass analyzer are detected by a
detector. In a time-of-flight mass analyzer, ions are accelerated
through a short high voltage field and drift into a high vacuum
chamber. At the far end of the high vacuum chamber, the accelerated
ions strike a sensitive detector surface at different times. Since
the time-of-flight of the ions is a function of the mass-to-charge
ratio of the ions, the elapsed time between ionization and impact
can be used to identify the presence or absence of molecules of
specific mass-to-charge ratio.
[0061] The time of flight data may then be converted into
mass-to-charge ratios to generate a spectrum showing the signal
strength of the markers as a function of mass-to-charge ratio. FIG.
2 shows a flowchart illustrating an exemplary method for converting
mass spectra based on time-of-flight data into mass-to-charge ratio
data. First, time of flight spectra are collected (step 16). Then,
a smoothing filter is applied to the time of flight spectra (step
18). Typically, a significant amount of high frequency noise is
present in the initially generated spectra. Various filters are
applied to reduce noise without corrupting the underlying signal.
Then, a baseline is calculated (step 20). This removes a
characteristic upward shift that can be characteristic of, for
example, a MALDI or a surface enhanced laser desorption/ionization
process.
[0062] "Surface enhanced" desorption/ionization processes refer to
those processes in which the substrate on which the sample is
presented to the energy source plays an active role in the
desorption/ionization process. In these methods, the substrate,
such as a probe, is not merely a passive stage for sample
presentation. Several types of surface enhanced substrates can be
employed in a surface enhanced desorption/ionization process. In
one example, the surface comprises an affinity material, such as
anion exchange groups or hydrophilic groups (e.g., silicon oxide),
that preferentially bind certain classes of molecules. Examples of
such affinity materials include, for example, silanol
(hydrophilic), C8 or C16 alkyl (hydrophobic), immobilized metal
chelate (coordinate covalent), anion or cation exchangers (ionic)
or antibodies (biospecific). The sample is exposed to a substrate
bound adsorbent so as to bind analyte molecules according to the
particular basis of attraction. Typcially non-binding molecules are
washed off. When the analytes are biomolecules, an energy absorbing
material, e.g., matrix, is typically associated with the bound
sample. Then a laser is used to desorb and ionize the analytes,
which are detected with a detector.
[0063] In another version, the substrate surface comprises a bound
layer of energy absorbing molecules, obviating the need to mix the
sample with a matrix material, as in MALDI. Surface enhanced
desorption/ionization methods are described in, e.g., U.S. Pat. No.
5,719,060 (Hutchens and Yip) and WO 98/59360 (Hutchens and Yip)
(U.S. Pat. No. 6,255,047). When a laser desorbs a matrix including
an energy absorbing material, some of the matrix material can also
be desorbed along with the sample material being analyzed. The
baseline calculation adjusts the spectra to take into account the
presence of the signal due to desorbed matrix material. Once a
baseline is calculated, a time of flight/mass transformation takes
place (step 22). In this step, the time of flight data is converted
into mass-to-charge ratios. Local noise values are then calculated
(step 24). At low mass-to-charge ratios, a significant amount of
noise is generated due to the desorbed matrix material. In an
ionization desorption process, desorption of the matrix material is
less likely at higher mass-to-charge ratios than at lower
mass-to-charge ratios. Noise is therefore more likely at lower
mass-to-charge ratios than at higher mass-to-charge ratios.
Adjustments to the spectra can be made to correct for this effect.
After these corrections are made, the spectra update is complete
(step 26). By processing mass spectra according to the method shown
in FIG. 2, the signal-to-noise ratio of the mass spectrum is
improved, allowing better quantitation and comparison of potential
markers.
[0064] Mass spectra data generated by the desorption and detection
of markers can be preprocessed using a digital computer after or
before generating a mass spectra plot. Data analysis can include
the steps of determining the signal strength (e.g., height of
signals) of a detected marker and removing "outliers" (data
deviating from a predetermined statistical distribution). For
example, the observed signals can be normalized. Normalization is a
process whereby the height of each signal relative to some
reference is calculated. For example, a reference can be background
noise generated by instrument and chemicals (e.g., an energy
absorbing molecule) which is set as zero in the scale. Then, the
signal strength detected for each marker or other substances can be
displayed in the form of relative intensities in the scale desired
(e.g., 100). Alternatively, a standard may be admitted with the
sample so that a signal from the standard can be used as a
reference to calculate relative intensities of the signals observed
for each marker or other markers detected.
[0065] The digital computer can transform the resulting data into
various formats for display. In one format, referred to as
"spectrum view or retentate map," a standard spectral view can be
displayed. The spectral view depicts the quantity of marker
reaching the detector at each particular molecular weight. In
another format, referred to as "peak map," only the peak height and
mass information are retained from the spectrum view, yielding a
cleaner image and enabling signals representing markers with nearly
identical molecular weights to be more easily seen. In yet another
format, referred to as "gel view," each mass from the peak view can
be converted into a grayscale image based on the height of each
peak, resulting in an appearance similar to bands on
electrophoretic gels. In yet another format, referred to as "3-D
overlays," several spectra can be overlaid to study subtle changes
in relative peak heights. In yet another format, referred to as a
"difference map view," two or more spectra can be compared,
conveniently highlighting signals representing markers and signals
representing markers that are up- or down-regulated between
samples. Marker profiles (spectra) from any two samples may be
compared visually on one plot. Data that can be used to form the
data set may be obtained from these and other mass spectra display
formats.
[0066] II. Forming the Data Set
[0067] Once the mass spectra are obtained, a data set such as a
known data set is formed. The data set comprises data that is
obtained from the mass spectra of the class set of biological
samples. The mass spectra data forming the data set can be raw,
unprocessed data. For example, raw signal intensity values at
identified mass values from the mass spectra may be used to form
the data set. In another example, raw signal patterns from mass
spectra may be used to form the data set.
[0068] In alternative embodiments, data may be preprocessed before
it is used to form the classification model. The mass spectra may
then be processed in any suitable manner before being used to form
the classification model. For example, the signals in the mass
spectra may be processed by taking the log values of the signal
intensities, removing outliers, removing signals which are less
likely to be associated with potential markers, removing signals
which have low intensities, etc.
[0069] In some embodiments, the data set may comprise raw or
preprocessed pattern data that relates to the particular pattern of
each mass spectrum. For example, for a mass spectrum comprising
many signal peaks, the pattern of the signal peaks may constitute a
fingerprint for the biological sample used to create the mass
spectrum. The classification process can classify the different
spectra according to patterns or pattern segments that may be
common to the spectra in the respectively different classes
differentiated by the classification model. A computer program such
as a neural network program, for example, can receive plural mass
spectra of known samples associated with known biological statuses.
The neural network can be trained with the mass spectra data so
that it can differentiate between mass spectra patterns belonging
to the respectively different classes. The trained neural network
can then be used to classify a mass spectrum associated with an
unknown sample based on the pattern in the mass spectrum.
[0070] In other embodiments, the data set comprises data relating
to the intensities of the signals in the mass spectra. In these
embodiments, some or all of the signals in each mass spectrum may
be used to form the data set. For example, the intensities of less
than all of the signals (e.g., peaks) in a spectra view type mass
spectrum can be used to form the data set. In preferred
embodiments, mass-to-charge ratios are identified, and the
identified mass-to-charge ratios are used to select signals from
the mass spectra. The intensities of these selected signals can be
used to form the data set. By using data from less than all signals
in each mass spectrum to form the data set, the number of data
points that will be processed is reduced so that data processing
occurs more rapidly. Data of signals that have a low likelihood of
representing acceptable markers may be excluded from the data
set.
[0071] Mass-to-charge ratios may be identified in any number of
ways. For example, the mass-to-charge ratios may be identified by
comparing the mass spectra of different classes having different
biological statuses. The mass-to-charge ratios of signals that are
likely to differentiate the classes may be selected. The comparison
may be performed manually (e.g., by a visual comparison) or may be
done automatically with a digital computer. For example, mass
spectra associated with different classes of samples can be
visually compared with each other to determine if the intensity of
a signal at a mass-to-charge ratio in a mass spectrum from one
sample class is significantly greater than or less than a signal at
the same mass-to-charge ratio in a mass spectrum from a different
sample class, thus indicating potential differential expression.
Mass-to-charge ratios where these signal differences occur may be
selected.
[0072] FIG. 3, for example, shows a graph of log (2) normalized
intensity vs. the identified peak clusters. This plot displays the
log base 2 normalized intensity values. Each intensity value in a
peak cluster has the average intensity value subtracted so a value
of zero represents no change from the average. Each unit on the
y-axis represents a two-fold difference from the cluster average.
Significantly up and down regulated proteins can be identified
using a plot such as the one shown in FIG. 3. FIG. 3 shows a graph
of log normalized intensity as a function of different signal
clusters. The signal intensities from mass spectra from two
different groups of samples are shown in the graph. For example,
the peak cluster 22 (on the x-axis) in FIG. 3 shows a wide
variation between the data points from Group A and Group B. This
indicates that the mass-to-charge ratio associated with peak
cluster 22 can be identified as a candidate marker location.
[0073] Alternatively or additionally, certain predefined criteria
may be provided to first select certain signals or signal clusters.
The selected signal clusters may then be used to identify
particular mass-to-charge ratios. For example, signals or signal
clusters having a signal intensity or average signal intensity
above or below a certain signal intensity threshold may be
automatically selected. Mass-to-charge ratios associated with these
selected signals or signal clusters may then be identified.
[0074] Preferred methods including collecting mass spectra data,
preprocessing the data, and processing the preprocessed mass
spectral data to form a classification model can be described with
reference to FIGS. 4 and 5. With reference to FIG. 4, mass spectra
of samples associated with different biological statuses are
collected (step 27). The number of samples collected is preferably
large. For example, in embodiments of the invention, the number of
collected samples may be from about 100 to about 1000 (or more or
less than these values). Preferably, all samples used to create the
spectra are created under similar conditions so that differences
between the samples are reflected in the spectra.
[0075] Signals corresponding to the presence of a potential marker
are identified in each spectrum. Each such signal is assigned a
mass value. Signals above a predetermined signal-to-noise ratio in
each mass spectrum in the first group of mass spectra are then
detected (step 28). In a typical example, signals with a
signal-to-noise ratio greater than a value S may be detected. The
value S may be an absolute or a relative value. Then, signals at
the mass-to-charge ratios in the mass spectra are clustered
together (step 30). Signal clusters that meet predetermined
criteria are then selected. For example, in one embodiment, signal
clusters having a predetermined number of signals can be selected
(step 32). Clusters having less than the predetermined number are
discarded. In a typical example, if the number of signals in a
cluster is less than 50% of the number of mass spectra, then the
signal cluster can be discarded. In some embodiments, the selection
process results in anywhere from as few as about 20 to more than
about 200 selected signal clusters. Once the signal clusters are
selected, the mass-to-charge ratios for these signal clusters can
be identified (step 34).
[0076] Once the mass-to-charge ratios are identified, "missing
signals" for the mass-to-charge ratios can be determined. Some of
the mass spectra may not exhibit a signal at the identified
mass-to-charge ratios. This group of mass spectra or the samples
associated with the mass spectra can be re-analyzed to determine if
signals do in fact exist at the identified mass-to-charge ratios
(step 36). Estimates are added for any missing signals (step 38).
For spectra where no signal is found in a cluster, an intensity
value is estimated from the trace height or noise value. The
estimated intensity value may be user selectable.
[0077] With reference to FIG. 5, once mass-to-charge ratios are
identified, intensity values are determined for each signal at the
identified mass values for all mass spectra (step 46). The
intensity value for each of the signals is normalized from 0 to 100
to remove the effects of absolute magnitude (step 48). Then, the
logarithm (e.g., base 2) is taken for each normalized signal
intensity (step 50). Taking the logarithm of the signal intensities
removes skew from the measurements.
[0078] The log normalized data set is then processed by a
classification process (step 52) that is embodied by code that is
executed by a digital computer. After the code is executed by the
digital computer, the classification model is formed (step 54).
Additional details about the formation of the classification model
are provided below.
[0079] III Forming the Classification Model
[0080] A classification process embodied by code that is executed
by a digital computer can process the data set The code can be
executed by the digital computer to create a classification model.
The code may be stored on any suitable computer readable media.
Examples of computer readable media include magnetic, electronic,
or optical disks, tapes, sticks, chips, etc. The code may also be
written in any suitable computer programming language including, C,
C++, etc.
[0081] The digital computer may be a micro, mini or large frame
computer using any standard or specialized operating system such as
a Windows.TM. based operating system. In other embodiments, the
digitial computer may simply be a one or more microprocessors The
digital computer may be physically separate from the mass
spectrometer used to create the mass spectra. Alternatively, the
digital computer may be coupled to or physically incorporated into
the mass spectrometer. Mass spectra data can be transmitted from
the mass spectrometer to the digital computer manually or
automatically. For example, in one embodiment, a known data set may
first be obtained from a plurality of mass spectra. The known data
set may then be manually entered into a digital computer running
code that embodies a classification process. In another embodiment,
the generation and/or collection of mass spectra data, the
preprocessing of the data, and the processing of the preprocessed
data by a classification process may be performed using the same
physical computational apparatus.
[0082] In some embodiments, the known data set can be characterized
as a training set which can "train" a precursor to the
classification model or a previously formed classification model.
The classification model may be trained and learn as it is formed.
For example, in a neural network, the known data set can be used to
train the neural network to recognize differences between the
classes of data that are entered into the neural network. After an
initial classification model is formed, a larger number of samples
can be used to further train and refine the classification model so
that it can more accurately discriminate between the classes used
to form the classification model.
[0083] In embodiments of the invention, additional data may be used
to form the classification model. The additional data may or may
not relate to mass spectra. For instance, in some embodiments,
pre-existing marker data may be used in addition to a known data
set to form the classification model. For example, mass spectra for
a class of prostate cancer patient samples and a class of
non-prostate cancer patient samples may be obtained. A known data
set may be formed using the mass spectra. A classification model
may be formed using the known data set and pre-existing marker data
such as pre-existing PSA diagnostic data (e.g., PSA clinical assay
data). The additional pre-existing PSA diagnostic data can be used
to help differentiate the mass spectra to form the classification
model. For example, each mass spectrum may be evaluated to see if a
signal at the mass-to-charge ratio corresponding to PSA is more
closely associated with a signal intensity characteristic of
prostate cancer or a signal intensity characteristic of
non-prostate cancer. This information can be used to help assign
the mass spectrum and its corresponding sample to a prostate cancer
or a non-prostate cancer class. In other embodiments, non-mass
spectra data such as the sex, age, etc. of the persons from which
the biological samples were taken may also be used to form a
classification model. For example, if men are more likely to have a
particular disease than women, then this information can also be
used to help classify samples and form a classification model.
[0084] Any suitable classification process may be used in
embodiments of the invention. For example, the classification
process may be a hierarchical classification process such as a
classification and regression tree process or a multivariate
statistical analysis. A multivariate statistical analysis looks at
patterns of relationships between several variables simultaneously.
Examples of multivariate statistical analyses include well known
processes such as discriminate function analysis and cluster
analysis. Discriminant function analysis is a statistical method of
assigning observations to groups based on previous observations
from each group. Cluster analysis is a method of analysis that
represents multivariate variation in data as a series of sets. In
biology, for example, the sets are often constructed in a
hierarchical manner and shown in the form of a tree-like diagram
called a dendrogram. Some types of cluster analyses and other
classification processes are described in the article by Jain et
al., "Statistical Pattern Recognition: A Review", IEEE Transactions
on Pattern Analysis and Machine Intelligence, Vol. 22, No. Jan. 1,
2000. This article is incorporated herein by reference in its
entirety.
[0085] Alternatively, the classification process may use a
non-linear classification process such as an artificial neural
network analysis. An artificial neural network analysis can be
trained using the known data set. In general, an artificial neural
network can predict the value of an output variable based on input
from several other input variables that can impact it. The
prediction is made by selecting from a set of known patterns the
one that appears most relevant in a particular situation. An
artificial neural network conceptually has several neuron elements
(units) and connections between them. These units are categorized
into three different layers or groups according to their functions.
A first group forms an input layer that receives the data entered
into the system. A second group forms an output layer that delivers
the output data representing an output pattern. A third group
comprises a number of intermediate layers, also known as hidden
layers that convert the input pattern into an output.
[0086] Illustratively, a neural network can be trained to
differentiate between laser desorption mass spectra associated with
a diseased state and a non-diseased state. Then, a mass spectrum of
a test biological sample can be created by a laser desorption
process and data relating to this mass spectrum can be input into
the trained neural network. The trained neural network can
determine if the test biological sample is associated with the
diseased state or non-diseased state.
[0087] In embodiments of the invention, the classification process
preferably includes a hierarchical, recursive partitioning process
such as a classification and regression tree process. In
embodiments of the invention, the classification and regression
tree process is embodied by computer code that can be executed by a
digital computer. An exemplary classification and regression tree
program is CART 4.0 commercially available from Salford Systems,
Inc. (www.salford-systems.com).
[0088] One specific classification and regression tree process is a
binary recursive partitioning process. The process is binary
because parent nodes are always split into exactly two child nodes
and recursive because the process can be repeated by treating each
child node as a parent. To partition a known data set, questions
are asked of the known data set. In embodiments of the invention,
the data being partitioned are the mass spectra corresponding to
the class set of biological samples. Each mass spectrum can be
considered an "instance" to be classified. An exemplary question
that may be used to partition the instances may be "Is the signal
intensity of the signal at the mass-to-charge ratio X greater than
Y?" Each question subdivides the known data set into two groups of
more homogeneous composition. Once a best split is found, the
classification and regression tree process repeats the search
process for each child node, continuing recursively until further
splitting is impossible or stopped. Splitting is impossible if only
one case remains in a particular node or if all the cases in that
node are of the same type.
[0089] The questions asked of the data set may be determined by a
user or may be automatically determined by a digital computer. In
some embodiments, the questions can be arbitrarily generated by a
digital computer and the quality of the data splitting determines
if the question is acceptable. For example, a question may be asked
of the data. If the partitioning results in a statistically
significant split of the instances, the question may be kept and
used to form the classification and regression tree. The
classification and regression tree process identifies the optimal
number of questions required to classify the data, compensating for
the effects of random error in each sample observation.
[0090] The classification and regression tree process looks at all
possible splits for all predictor variables included in the
analysis. For example, for a data set with 215 instances and 19
predictor variables, the process considers up to 215 times 19
splits for a total of 4085 possible splits. Typically, all such
splits are considered when forming a classification and regression
tree. Consequently, the formed classification and regression tree
process takes into account many different predictor variables in
forming the classification model. For example, in a typical
embodiment, data of signals at over 100 mass-to-charge ratios in
all mass spectra for the class set are taken into account when
forming the classification model. In comparison, the differential
expression analysis described above takes only one predictor
variable into account. Consequently, the classification and
regression tree embodiments can provide more accurate
classification accuracy than other classification methods since
more data from each mass spectrum is used to form the
classification model.
[0091] To check the accuracy of the model, the classification and
regression tree process may employ a computer-intensive technique
called cross validation. In a typical cross-validation process, a
large tree is grown and is then pruned back. The data set is
divided into 10 roughly equal parts, each containing a similar
distribution for the biological statuses being analyzed. The first
9 parts of the data are used to construct the largest possible
tree. The remaining 1 part of data is used to obtain initial
estimates of the error rate of selected sub-trees. The same process
is then repeated (growing the largest possible tree) on another
{fraction (9/10)} of the data while using a different {fraction
(1/10)} part as the test sample. The process continues until each
part of the data has been held in reserve one time as a test
sample. The results of the 10 mini-test samples are then combined
to form error rates for trees of each possible size. These error
rates are applied to the tree based on the entire data set. Cross
validation provides fairly reliable estimates of the independent
predictive accuracy of the tree. Even if an independent test sample
is not available, a prediction can be made as to how accurately the
tree can classify completely fresh data (e.g., data from a
plurality of unknown samples).
[0092] The classification and regression tree that is created
provides a representation of which of the predictor variables (if
any) are responsible for the differences between sample groups. The
classification and regression tree can be used for classification
(predicting what group a case belongs to) and also be used for
regression (predicting a specific value). It can also be used to
identify features that may be important in discriminating between
the classes being analyzed. For example, the classification model
may indicate that one or more signal intensity values at specific
mass-to-charge ratios, alone or in combination, are important
features that differentiate the classes being analyzed.
[0093] The classification and regression tree graphically displays
the relationships found in data. One primary output of the
classification and regression tree process is the tree itself. The
tree can serve as one aspect of a classification model that can be
visually analyzed by a user. Unlike non-linear techniques such as a
neural network analysis, the visual presentation provided by the
tree makes the classification analysis very easy to understand and
assimilate. As a result, users tend to trust the results of
decision trees more than they do "black box" classification models
such as those characteristic of trained neural networks. This makes
the classification and regression tree a desirable classification
model for various health care and regulatory personnel (e.g., the
Food and Drug Administration), and patients, who may want to have a
detailed understanding of the analysis used to create the
classification model. The trees can also be used to discover
previously unknown connections between the data and the biological
statuses being analyzed.
[0094] The classification and regression tree process has other
advantages over classification processes such as a neural network
analysis. For example, classification and regression tree programs
are more efficient than neural networks, which typically require a
large number of passes of the training set data, sometimes
numbering in the thousands. The number of passes required to build
a decision tree, however, is no more than the number of levels in
the tree. There is no predetermined limit to the number of levels
in the tree, although the complexity of the tree as measured by the
depth and breadth of the tree generally increases as the number of
predictor variables increases.
[0095] Also, using the classification and regression tree model,
features that may discriminate between the classes may be
identified. The identified features in the data may be
characteristic of the biological status(s) being analyzed. For
example, the classification model may indicate that a combination
of features is associated with a particular biological status. For
example, the model may indicate that specific signal intensities at
different mass-to-charge ratios differentiate a diseased state from
a non-diseased state. In comparison to conventional differential
analysis processes, in embodiments of the invention, many different
variables may be analyzed. The classification model can identify a
single predictor variable or can identify multiple predictor
variables that may differentiate the biological statuses being
analyzed.
[0096] IV. Using the Classification Model
[0097] The classification model may be used to classify an unknown
sample into a biological status. In this method the mass spectrum
of a test sample can be compared to the classification model
associated with a particular biological status to determine whether
the sample can be properly classified with the biological status. A
mass spectrum of the unknown biological sample can be obtained, and
data obtained from a mass spectrum of the unknown sample can be
entered into a digital computer. The entered data may be processed
using a classification model. The classification model may then
classify the unknown sample into a particular class. The class may
have a particular biological status associated with it, and the
person can be diagnosed as having that particular biological
status.
[0098] This method has particular use for clinical applications.
For example, in the process of drug discovery, one may wish to
determine whether a candidate molecule produces the same
physiological result as a particular drug or class of drugs (e.g.,
the class of seratonin re-uptake inhibitors) in a biological
system. A classification model is first developed that
discriminates biological systems based on exposure to the drug or
class of drugs of interest (e.g., persons or test animals). Then,
the biological system is exposed to the test molecule and a mass
spectrum of a sample from the system is produced. This spectrum is
then classified as belonging or not belonging to the classification
of known drug or group of drugs against which it is being tested.
If the candidate molecule is assigned to the class, this
information is useful in determining whether to perform further
research on the drug.
[0099] In another application, a classification model is developed
that discriminates various toxic and non-toxic biological states.
Toxic status can result from, e.g., exposure to a drug or class of
drugs. That is, a classification model can be developed that
indicates whether or not a drug or class of drugs produces a toxic
response in a biological system (e.g., in vivo or in vitro model
systems including liver toxicity). Then, a drug that is in
development or in clinical trials can be tested on the system to
determine whether a spectrum from a sample from the system can be
classified as toxic or not. This information also is useful in
toxicity studies during drug development.
[0100] In another application, a classification model is developed
that discriminates between persons who are responders and
non-responders to a particular drug. Then, before giving a drug to
a person who is not known to be a responder or non-responder, a
sample from the person is tested by mass spectrometry and assigned
to the class of responders or non-responders to the drug.
[0101] In another application, a classification model is developed
that distinguishes person having a disease from those who do not
have the disease. Then a person undergoing diagnostic testing can
submit a sample for classification into the status of having the
disease and not having the disease. Thus, this method is useful for
clinical diagnostics.
[0102] One embodiment is directed to analyzing cancer. Pathologists
grade cancers according to their histologic appearance. Features of
low-grade cancers include enlarged nuclei with a moderate increase
in nuclear/cytoplasmic ratio, small number of mitoses, moderate
cytologic heterogeneity, and retention of generally normal
architecture. Features of high-grade cancers include enlarged,
bizarre looking nuclei with a high nuclear/cytoplasmic ratio;
increased number of mitoses, some of which may appear atypical; and
little or no resemblance to normal architecture. It is useful to
develop a classification model that distinguishes a biological
sample coming from un-diseased, low-grade cancer, and high-grade
cancer, since this diagnosis often dictates therapeutic decisions
as well as can predict prognosis. The sample can be a solid tissue
biopsy or a fine needle aspirate of the suspected lesion. However,
in another embodiment, the samples can derive from more easily
collected sources from the group of individuals being tested, such
as urine, blood or another body fluid. This is particularly useful
for cancers that secrete cells or proteins into these fluids, such
as bladder cancer, prostate cancer and breast cancer. Upon
establishment of the classification model for these states, the
model can be used to classify a sample from a person subject to
diagnostic testing. In another application, a classification model
is developed that discriminates between classes of individuals
having a particular physical or physiological trait that is not
pathologic. Then, individuals unknown to have the trait can be
classified by testing a sample from the individual and classifying
a spectrum into the class having the trait, or outside the class
having the trait.
[0103] The classification model can also be used to estimate the
likelihood that an unknown sample is accurately classified as
belonging to a class characterized by a biological status. For
instance, in a classification and regression tree, the likelihood
of potential misclassification can be determined. Illustratively, a
classification and regression tree model that differentiates a
diseased state from a non-diseased state classifies an unknown
sample from a patient. The model can estimate the likelihood of
misclassification. If, for example, the likelihood of disease
misclassification is less than 10%, then the patient can be
informed that there is a 90% chance that he has the disease.
[0104] V. Systems Including Computer Readable Media
[0105] Some embodiments of the invention are directed to systems
including a computer readable medium. A block diagram of an
exemplary system incorporating a computer readable medium and a
digital computer is shown in PIG. 6. The system 70 includes a mass
spectrometer 72 coupled to a digital computer 74. A display 76 such
as a video display and a computer readable medium 78 may be
operationally coupled to the digital computer 74. The display 76
may be used for displaying output produced by the digital computer
74. The computer readable medium 78 may be used for storing
instructions to be executed by the digital computer 74.
[0106] The mass spectrometer can be operably associated with the
digital computer 74 without being physically or electrically
coupled to the digital computer 74. For example, data from the mass
spectrometer could be obtained (as described above) and then the
data may be manually or automatically entered into the digital
computer 74 using a human operator. In other embodiments, the mass
spectrometer 72 can automatically send data to the digital computer
74 where it can be processed. For example, the mass spectrometer 72
can produce raw data (e.g., time-of-flight data) from one or more
biological samples. The data may then be sent to the digital
computer 74 where it may be pre-processed or processed.
Instructions for processing the data may be obtained from the
computer readable medium 78. After the data from the mass
spectrometer is processed, an output may be produced and displayed
on the display 76.
[0107] The computer readable medium 78 may contain any suitable
instructions for processing the data from the mass spectrometer 72.
For example, the computer readable medium 78 may include computer
code for entering data obtained from a mass spectrum of an unknown
biological sample into the digital computer 74. The data may then
be processed using a classification model. The classification model
may estimate the likelihood that the unknown sample is accurately
classified into a class characterized by a biological status.
[0108] Although the block diagram shows the mass spectrometer 72,
digital computer 74, display 76, and computer readable medium 78 in
separate blocks, it is understood that one or more of these
components may be present in the same or different housings. For
example, in some embodiments, the digital computer 74 and the
computer readable medium 76 may be present in the same housing,
while the mass spectrometer 72 and the display 76 are in different
housings. In yet other embodiments, all of the components 72, 74,
76, 78 could be formed into a single unit.
EXAMPLE
[0109] A plurality of mass spectra was generated from biological
samples from a set of biological samples. The set included a first
class of serum from normal patients and a second class of serum
from patients with prostate cancer. A serum sample from each
patient was run through a surface enhanced laser
desorption/ionization system commercially available from Ciphergen
Biosystems, Inc. of Fremont, Calif. Ciphergen Biosystem's
ProteinChip.RTM. technology was also used in this example.
Additional details about ProteinChip.RTM. technology can be found
at the Website www.ciphergen.com. The resulting output for each
sample was a mass spectrum plot of signal intensity vs.
mass-to-charge ratio. Discrete peaks represented the signals in the
mass spectra.
[0110] The intensities of the signals at the particular
mass-to-charge ratios corresponded to the amount of proteins having
the particular mass-to-charge ratios. For example, high signal
intensities indicate high concentrations of proteins. Signals in
each mass spectrum were located, quantified, and selected. In this
example, segments of a mass spectrum were considered acceptable
signals if they had intensity values at least twice as great as the
surrounding noise level. Signals in the mass spectra at
approximately the same mass-to-charge ratios were clustered
together in all mass spectra. After clustering, about 250 signal
clusters were identified and were labeled P1 through P250. Each
signal cluster, P1 through P250, corresponded to a specific
mass-to-charge ratio and was characterized as a "predictor
variable".
[0111] The signal intensities at the identified mass-to-charge
ratios for each mass spectrum formed the known data set. These
signal intensities were entered into a classification and
regression tree program, CART 4.0, commercially available from
Salford Systems, Inc. (www.salford-systems.co- m). The program was
executed by a digital computer. The digital computer formed a
classification and regression tree. Using the data, each sample was
classified as normal or cancer.
[0112] After the mass spectra data was input, the digital computer
produced a tree such as the one shown in FIG. 6. In this example,
class 0 is normal while class 1 is cancer. Each mass spectrum can
be characterized as an "instance" which is classified in the
tree.
[0113] Each box in the tree represents a "node". The top node, Node
1, is called the root node. The decision tree grows from the root
node, splitting the data at each level to form new nodes. Branches
connect the new nodes. Nodes that do not experience further
splitting are called terminal nodes. The terminal nodes in the tree
shown in FIG. 6 are labeled Terminal Nodes 1 to 7. As will be
explained in further detail below, Terminal Nodes 1 to 7 can be
used to classify an unknown sample and can thus be used for
prediction.
[0114] In each node, the majority sets the classification for the
entire node. For example, Terminal Node 1 has four patients. Of
these four patients, all four patients have cancer. Terminal Node 1
is therefore characterized as a cancer node. Because all instances
have the same value (cancer), this node is characterized as "pure"
and will not be split further. If Terminal Node 1 included three
cancer patients and one normal patient, the node would still be
characterized as a cancer node since a majority of the patients are
cancer patients. In this example, the one normal patient would be
considered incorrectly classified.
[0115] In FIG. 6, each node contains information about the number
of instances at that node, and about the distribution of the
biological status, cancer. The instances at the root node (Node 1)
are all of the instances in the mass spectra data set. Node 1
contains 194 instances, of which 96 are normal and 98 are cancer.
Node 1 is splits into two new nodes, Node 2 and Node 5. The data
split is determined by determining whether the average signal
intensity for the cluster P127 is less than or equal to 3.2946. The
average signal intensities, as well as the value 3.2946 were on a
relative scale. If the answer to this question is yes, then the
corresponding instances are placed in Node 2. If the answer to this
question is no, then the corresponding instances are placed in Node
5. In this example, the mass spectra of 85 cancer patients and 11
normal patients had a signal intensity less than or equal to 3.2946
at the mass-to-charge ratio associated with the predictor variable
P127 and were placed in Node 2. The mass spectra of 85 normal
patients and 13 cancer patients had a signal intensity greater than
3.2946 at the mass-to-charge ratio associated with the predictor
variable P127 and were placed in Node 5. Similar partitioning using
different splitting rules occurred at the other nodes to form the
tree.
[0116] The prediction performance of the classification and
regression tree can be described with reference to the Tables 1 and
2.
1TABLE 1 Misclassification for Learn Data Class N Cases N
Misclassified Percent Error 0 (Normal) 96 0 0 1 (Cancer) 98 0 0
[0117]
2TABLE 2 Misclassification for Test Data Class N Cases N
Misclassified Percent Error 0 (Normal) 96 9 9.38 1 (Cancer) 98 11
11.22
[0118] The classification and regression tree program divided the
known data set into two groups. About 90% of the data was used as a
learning set and about 10% was used as a test set. A classification
and regression tree is initially formed using the learning set
data. After the tree was formed, it was tested with the remaining
10% test data to see how accurately the classification and
regression tree classifies data. With reference to Table 1, all of
the learning set data was corrected classified using the formed
classification and regression tree. With reference to Table 2, the
percent error rates for classifying the normal case and the cancer
case test data were 9.38% and 11.22%, respectively. Conversely, the
classification success rate was 90.62% and 88.78% for the normal
cases and the cancer cases, respectively.
[0119] Classification success rates such as these indicate that the
classification and regression tree is a highly accurate model for
classifying unknown biological samples. In the classification
process, multiple predictor variables are considered in the
classification scheme. Much more data can be used from a mass
spectrum to classify the sample associated with the mass spectrum
than the previously described differential analysis procedure,
which only uses average signal intensities at a single
mass-to-charge ratio to classify a test patient. Accordingly, the
classification model can be more accurate in classifying a test
patient then many conventional classification models.
[0120] Once grown, the tree can be used to classify an unknown
sample by starting at the root (top) of the tree and following a
path down the branches until a terminal node is encountered. The
path is determined by imposing the split rules on the values of the
predictor variables in the mass spectrum for the unknown sample.
For example, if a mass spectrum of an unknown serum sample from a
test patient has signals with intensities of 1.0, 0.05, and 0.9 at
the mass-to-charge ratios of predictor variables P127, P193, and
P187 respectively, then the test patient would be classified in
Node 1, Node 2, Node 3, and then finally Terminal Node 1. Terminal
Node 1 is a cancer node and the patient would be classified as
being a cancer patient.
[0121] FIG. 7 shows a table of variable importance of each of some
of the predictor variables (e.g., signal clusters). The variable
importance table ranks the predictor variables by how useful they
were in building the classification and regression tree. If a
specific predictor variable strongly differentiates the mass
spectra data, then it is important in building the classification
tree. To calculate a variable importance score, CART looks at the
improvement measure attributable to each variable in its role as a
surrogate to a primary split. The values of these improvements are
summed over each node and totaled, and are scaled relative to the
best performing variable. The variable with the highest sum of
improvements is scored 100, and all other variables will have a
lower score ranging downwards towards zero.
[0122] In FIG. 7, the classification model indicates that the
predictor variables P36, P127, and P90 are more important than
other predictor variables in forming the classification and
regression tree. They are consequently more important than other
predictor variables in discriminating between the classes, cancer
and non-cancer. The mass-to-charge ratios associated with these
predictor variables are also associated with potential markers that
differentiate prostate cancer samples from non-prostate cancer
samples. Accordingly, the classification model can be used to
identify one or more markers that may discriminate between classes
being analyzed.
[0123] The effectiveness of the tree model can be confirmed with
reference to FIGS. 8 and 9. The views in FIG. 8 are gel views while
the views in FIG. 9 are trace views. The spectra are zoomed into
the signal represented by P127 at a mass-to-charge ratio of 5075
daltons (charge=+1). FIGS. 8 and 9 show that markers in samples
from six prostate cancer patients and six non-prostate cancer
patients are differentially expressed at the mass value of 5075
daltons corresponding to the predictor variable P127. As shown in
the tree in FIG. 6, the predictor variable P127 is the first node
in the tree. Also, as shown in FIG. 7, the predictor variable P127
was shown to be more effective in differentiating the prostate
cancer class of samples from the non-prostate cancer patient class
of samples than most other predictor variables.
[0124] While the foregoing is directed to certain preferred
embodiments of the present invention, other and further embodiments
of the invention may be devised without departing from the basic
scope of the invention. Such alternative embodiments are intended
to be included within the scope of the present invention. Moreover,
the features of one or more embodiments of the invention may be
combined with one or more features of other embodiments of the
invention without departing from the scope of the invention.
[0125] All publications (e.g., Websites) and patent documents cited
in this application are incorporated by reference in their entirety
for all purposes to the same extent as if each individual
publication or patent document were so individually denoted. By
their citation of various references in this document Applicants do
not admit that any particular reference is "prior art" to their
invention.
* * * * *
References